hanamizuki.tw

OMG, Threads now auto-translates posts into your language. A while back I noticed my feed filling up with Korean, Japanese, and languages I can't read at all. Clearly posts I'd be into, but I couldn't read a word. Now they show up translated. A feed with no borders.

"What matters for the long term, and is going beyond all of these cycles of who is getting rich in one year, who is not getting rich in one year, is the underlying goals." @VitalikButerin , on stage at @ETHTaipei an hour ago. I expected the talk to be dense protocol math.

solopreneur 0.6.1 is out! The best part of this release is what's gone: 1,001 lines of skills that taught my agents things that are outdated. Also in 0.6.1: ➊ /preview now renders proposals to a local library by default, vercel publishing is opt-in.

I knew onboarding was supposed to matter. Pulling the actual numbers still surprised me. 55% of everyone who has ever paid for my app @glp1_tracker paid on the day they installed it, most of them within five minutes. I redid the onboarding flow in July and that went up to 70%.

I'm honestly not sure this one is a problem. People who don't buy on day 0 mostly don't come back the next day either (day 1 and day 2 add up to 7.3%), and right now I do nothing to bring them back. But it might just be that if someone wasn't convinced in the moment, they were never going to be.

I knew onboarding was supposed to matter. Pulling the actual numbers still surprised me. 55% of everyone who has ever paid for my app @glp1_tracker paid on the day they installed it, most of them within five minutes. I redid the onboarding flow in July and that went up to 70%.

Things I want to test: which plan is preselected, what order the plans show up in, how the countdown looks, and whether the full-price option is worth keeping at all. Second, the early-bird price and the deadline.

I knew onboarding was supposed to matter. Pulling the actual numbers still surprised me. 55% of everyone who has ever paid for my app @glp1_tracker paid on the day they installed it, most of them within five minutes. I redid the onboarding flow in July and that went up to 70%.

88% of buyers take the early-bird price. So that's not really a discount, that's just my price. The full price is sitting there as an anchor. So here's what I want to poke at, roughly in the order of how much I think it matters. First, onboarding.

summer.fi got drained for $6M. fact is, AI is making exploits cheaper. AI is making exploits faster. and most DeFi teams can barely keep the lights on. not enough users. cutting headcount. security gets deprioritized. bad cycle. getting worse. so i built this.

then i added alerts. 12 security accounts monitored around the clock — PeckShield, CertiK, SlowMist, zachxbt. bell any protocol. something goes wrong, you get an email. not 47 separate emails. one clean summary per hour.

summer.fi got drained for $6M. fact is, AI is making exploits cheaper. AI is making exploits faster. and most DeFi teams can barely keep the lights on. not enough users. cutting headcount. security gets deprioritized. bad cycle. getting worse. so i built this.

isthissafetoape.com paste any DeFi URL. security score from 0 to 100 across six dimensions: • smart contracts • tokenomics • governance • team background • market position • infrastructure no audit? hard cap at 60. anon team + no multisig? cap at 50.

Day 18 of building innieai.app. Idea mining for creators. The problem with backend scrapers is they don't see what you see. (Login walls, dynamic JS, paywalls, Bot blockers...) Today, I built a client-side extraction in both Chrome Extension and iOS Share Extension. 👇

Now, when you share a link from a browser, innie captures the actual content you are looking at. There are some nuances. As innie is cross-platform (ios/android/web), Some platforms doesn't support this method. 👇

Day 17 of building innieai.app, a black cat lives in your bookmarks. Today was all about UX polishing. If the chat feels clunky, the AI's intelligence doesn't matter. ✅ Markdown Rendering ✅ Code Block Copy ✅ Unified Logging The gap between "Prototype" and "Product" is in these details. 🧵

Behind the scenes: • Fixed Pagination (Total counts were lying to me 🤥). • Unified Logging: Replaced 17 scattered loggers with a single, clean extension. • Backend Stability: Increased Uvicorn workers & timeouts. No more hanging requests.

Day 17 of building innieai.app, a black cat lives in your bookmarks. Today was all about UX polishing. If the chat feels clunky, the AI's intelligence doesn't matter. ✅ Markdown Rendering ✅ Code Block Copy ✅ Unified Logging The gap between "Prototype" and "Product" is in these details. 🧵

Day 16 of building innieai.app. Helping thinkers become creators. Inspired by Claude Code, I wanted users to see how the AI thinks. So today I implemented Visible Thinking Process. No more staring at a blank screen waiting for a reply. Now you see the brain turning. ⚙️

The Tech Stack Backend: • Wraps internal monologue in <think> tags. • Streams via SSE: {"type": "thinking", "content": "Analyzing user intent..."} Frontend (iOS/Android/Web): • Renders a collapsible "Thinking..." block. • Updates in real-time as the stream arrives.

Day 15 of building innieai.app who interviews you about what you (might) read. Today I built the memory system using Neo4j. Now innie remembers your preferences, not just the current chat window. Here is the architecture behind the "Memory" vs "Diary" split. 🧵

I split user data into two nodes: 1️⃣ Memory (Synthesized): Stores traits ("Likes coffee"). and merges old & new info. (e.g., "Likes coffee" -> "Switched to Tea") using Gemini Flash subagent. 2️⃣ Diary (Append-only): Stores events ("Went to Tokyo"). Direct append. No LLM cost.

Day 14 of building innieai.app. Turning bookmarks into thoughts. Today I refactored the chat loop into a ReAct Architecture, so innie graduated from a "Chatbot" to an "Agent." Now, instead of just guessing, innie can use tools to find the answer. 🛠️

Handling tool outputs in the stream was tricky. • Designed a cleaner protocol: {"type": "tool_start", "label": "Searching Signals..."} • Frontend now shows cool animations when tools are running, so users know what Innie is doing. Transparency builds trust.

Day 14 of building innieai.app. Turning bookmarks into thoughts. Today I refactored the chat loop into a ReAct Architecture, so innie graduated from a "Chatbot" to an "Agent." Now, instead of just guessing, innie can use tools to find the answer. 🛠️

Day 12 building innieai.app The app that proactively asks you questions about your thoughts. The hardest thing for engineers to learn: build less features. innie is built on a graph database, and I almost added RELATED_TO between Topics. Claude Code ↔ LLM ↔ AI ↔ Productivity...

Was even thinking about adding weights. Then I asked myself: "...where do these relationships come from?" and I figured, they're already in the data. When an article is tagged with both "Claude Code" and "Productivity", those two concepts have already co-occurred.

Been playing with xAI Grok API lately. Found that the new tool "x_search" can just let you fetch X posts directly. Used to need X API subscription to fetch posts. Now just buy some xAI credits. But it's too new, Claude only knows Grok 2 stuff. and we are on Grok 4.1 now,

Day 11 of building @innie_app. I killed my AI Summary Bot today. I am a huge fan of @n8n_io. I built a workflow that read my bookmarks daily, and mailed me summaries in Ronny Chieng's tone from The Daily Show. In the beginning I loved it, but after a while, I never opened them.

Summaries are passive. They don't make you think. So, I'm trying something different: The "Worth Asking" Algorithm. Instead of summarizing everything, innie filters for Conversational Potential. I spent the day tuning a new logic: 1. Cluster: Group recent bookmarks via embeddings.